Machine Learning Papers

AI agents increasingly act as analysts whose decision can often be cast as hypothesis tests. Classical tests control the Type I error rate, the probability of rejecting a true null hypothesis, for fixed procedures. However, AI analysts introduce internal randomness and can evolve continuously over time. As AI analysts evolve, the risk of approving versions with Type I error rates above the target can accumulate, and adaptive use of evaluation feedback can lead to overfitting. We introduce continual validation, which evaluates each proposed AI analyst version before deployment using simulated data under the null hypothesis, with scenarios that can be randomized and expanded over time. We prove that, regardless of how updates are chosen or how many are proposed, the lifetime probability of ever deploying a version whose Type I error rate exceeds the allowed level remains bounded. This guarantees that every subsequent decision controls the Type I error rate at the target level across these scenarios. At the same time, updates with sufficiently low Type I error rates pass validation with high probability, allowing the analyst to continuously improve its performance without sacrificing rigor. In two examples involving evolving AI analysts for bioequivalence studies and product A/B tests, Type I error rates reached 0.098 and 0.183 without continual validation, compared with at most 0.045 and 0.044 with it, while retaining 82% and 96% of the power from optimally selected valid updates.

We give an algorithm for computing an $\eps$-approximate fixed point of any $d$-dimensional nonexpansion with respect to the $\ell_\infty$ norm in time $\log(1/\eps)^{O(\log^2 d)}$. At its core, our method uses a boosting argument to geometrically shrink a search region. Among other consequences, our algorithm implies that simple stochastic games can be solved in quasipolynomial time.

Diffusion and flow models provide expressive policy classes for robotic and sequential decision-making systems. However, their action likelihoods are generally intractable, limiting the direct use of widely adopted policy optimization methods such as PPO and GRPO. We propose Doob Policy Optimization (DoobPO), a framework that leverages Doob's $h$-transform for likelihood-free reinforcement learning fine-tuning of generative policies. DoobPO separates policy improvement from generative policy learning. It first learns a lightweight ratio model that represents the desired policy update, and then distills the resulting ratio-tilted target policy into the generative policy via conditional drift matching. This formulation applies uniformly to both diffusion and flow policies and supports a broad family of policy improvement rules, including PPO and GRPO. We prove that the practical conditional drift matching objective is gradient-equivalent to the ideal Doob drift matching objective. We further show that the induced diffusion process solves an endpoint-constrained stochastic optimal control problem. This establishes that it achieves the minimum-energy modification required to realize the desired policy tilt. Across MuJoCo continuous-control benchmarks, DoobPO achieves strong performance against competitive diffusion- and flow-policy baselines, showing a principled and practical route to RL fine-tuning for generative policies.

Program optimization aims to improve performance while preserving behavior. Coding agents can already make programs faster by changing algorithms, data structures, and execution strategies. However, existing equivalence checks mostly rely on tests, which agents have been shown to exploit, passing the tests without preserving the intended behavior. We propose agentic verified optimization, in which an agent produces a faster implementation together with a machine-checked proof of equivalence to the existing one. Because the existing implementation serves as the specification, the task avoids the specification-writing bottleneck of verified code generation, and recent advances in theorem proving make the proofs attainable. We formulate the task at the function and repository levels in Lean 4, where programs can be both executed and reasoned about. Few real-world applications are written in Lean, so we develop Aqueduct, an agentic framework that migrates C, C++, Go, and Rust programs into native Lean, validates each port by differential testing against its original. On the 189 reference implementations in Verina, agents produce faster proved rewrites for 156, including algorithmic changes such as replacing exponential recursion with dynamic programming. On 17 migrated applications, their 92 proved rewrites speed up 47 of 51 application scenarios. These application-level gains are smaller than the function-level ones, making the choice of functions to optimize an important problem. Current agents can thus already carry out verified optimization on both standalone functions and applications migrated from real programs. We discuss the remaining challenges and the steps toward verified optimization in production.

Speech models are increasingly used as an interface to AI systems, from transcription tools to voice agents. Existing audio backdoors require an attacker to introduce triggers at inference time. We study a different threat model in which the attacker compromises the model during fine-tuning but has no access to the victim's inference-time inputs. We introduce the Speaker-Activated Backdoor (SAB) setting, in which a target speaker's natural voice serves as the trigger. We develop a two-phase fine-tuning method that first makes speaker identity more accessible and then binds the target speaker to an attacker-chosen payload. Compared with a one-phase baseline with the same training budget, the two-phase method substantially improves selectivity and clean performance, and is especially important for content-dependent attacks. SAB achieves 90.2% and 97.0% backdoor success on Whisper-medium and Parakeet, with false-positive rates of 1.7% and 0%, respectively using as little as two minutes of real audio from the target speaker. We further demonstrate content-dependent semantic rewriting and downstream tool-use attacks, where malicious instructions are inserted into 96.9% of target-speaker transcripts. When these transcripts are passed to frontier LLM agents, GPT-4o and Claude (Opus, Sonnet, and Haiku), the injected instruction drives the agent to carry out the harmful action in up to 87.5% of cases. The attack generalizes to held-out recordings, multiple target speakers, and common acoustic degradations. These results establish speaker identity as a practical backdoor trigger that requires no inference-time attacker access and is invisible at the transcript layer.

The success of neural networks is largely attributed to their ability to learn useful representations from training data. However, known results that prove a computational advantage of learned features over fixed ones are restricted to specific input distributions. We settle the COLT open question of Feldman, Kamath, and Srebro (2026) by showing that this advantage can persist in the distribution-free setting. We construct a family of hypothesis classes that can be efficiently learned by a natural gradient-based algorithm on a two-layer ReLU network, but cannot be represented by linear predictors over any fixed feature map with polynomial dimension (and runtime). Specifically, for any sufficiently small error and any sufficiently large runtime budget R (depending on the error), there exists a hypothesis class that the neural network learns under any input distribution to the desired expected classification error in time at most R, while the dimension complexity (governing the dimension needed by linear predictors) is superpolynomial in R.

Uniform convergence asks how many random samples suffice for empirical averages to approximate expectations simultaneously over an entire class of functions. It is a central concept in probability and a foundation of statistical learning. For binary-valued classes, the Vapnik--Chervonenkis theorem characterizes uniform convergence by finiteness of the VC dimension. For dimension $d$, the sharp worst-case sample complexity for additive error $\varepsilon$ is of order $d/\varepsilon^2$. For real-valued functions, the analogous combinatorial parameter is the fat-shattering dimension, which depends on an approximation scale $γ$. The challenge is to achieve error $γ+\varepsilon$ with order $d/\varepsilon^2$ samples, matching the sample bound for binary classes, where $d$ is now the fat-shattering dimension at scale $γ$. Previous results achieved either the optimal scale or the optimal sample complexity, but not both. Our main result closes this gap, establishing the sharp sample complexity at the optimal scale. Along the way, we use several tools that may be of independent interest. The first concerns $\varepsilon$-nets with margin. A classical $\varepsilon$-net intersects every set of mass at least $\varepsilon$ in a prescribed family; we allow the net to intersect a slightly enlarged version of each such set. Remarkably, this relaxation permits bounds independent of the ambient dimension in natural geometric settings. Second, we derive improved covering bounds in both discrete and real-valued settings: we extend Haussler's covering lemma to partial concept classes and use this extension to obtain $\ell_2$ covers at the optimal scale for real-valued function classes. Third, we obtain a chaining bound at the optimal scale by aligning the covers so that successive approximations become increasingly close to one another.

Recent research-level physics-reasoning benchmarks for LLMs reveal a failure mode that lies not in knowledge or execution, but in modeling the world being reasoned about. Models often silently alter given quantities, fabricate laws, or inject unstated assumptions to force a problem into familiar patterns, and then reason inside this corrupted world. We define this premise-level divergence as modeling drift: the gap between a model's internal axiomatic world model (AWM) and the true physical world specified in the problem. Although RLVR and PRM have yielded substantial gains in mathematics and code, their reward shapes do not directly supervise world modeling in axiomatic domains such as physics. We propose Reinforcement Learning with Axiomatic World Modeling (RLAWM), which optimizes the policy by penalizing divergence between its inferred axiomatic world model and the true physics world. RLAWM probes physics reasoning at two hierarchical levels (modeling and solving), each scored by a physically grounded reward. The modeling reward targets abstract knowledge formulation, evaluating whether the model captures the true axiomatic world before attempting calculation. The solving reward targets instantiated knowledge reasoning, checking whether the model's concrete derivation remains consistent with the axiomatic world. This two-level structure separates what world the model believes it is reasoning in from how it reasons within that world. The framework operates exclusively during training, with no extra cost during inference. RLAWM improves performance by up to +12.2 points on PhysReason and +5.2 on PHYSICS over the baseline. It also generalizes zero-shot across difficulty, source, and domain, with a 2.45x gain and +10 points over the strongest baseline on olympiad-level physics and math.

Controllable 3D part generation is pivotal for digital asset creation, yet existing methods lack the flexibility to accommodate diverse user workflows requiring varying levels of control. In this paper, we introduce FlexPart, a unified image-to-3D part generation framework that supports variable-granularity geometric conditions within a single model. By seamlessly integrating heterogeneous inputs, ranging from sparse points to dense bounding boxes and masks, via gated adaptive modulation, FlexPart unlocks a powerful cross-modal synergy. We also propose an asymmetric geometric guidance mechanism that leverages strong spatial priors to align representations across granularities, improving geometric consistency and facilitating inter-task synergy. Under mask conditioning on PartVerseXL, FlexPart reduces Part CD by 16.6% relative to TripoSG + SAM3D and by 10.5% relative to OmniPart, while maintaining competitive whole-mesh quality. The code can be found at https://github.com/Gucci233/FlexPart

We introduce Foundation4D, a unified feed-forward model for geometry estimation, 3D detection, semantic occupancy, and Gaussian reconstruction from multi-camera driving video. A shared backbone uses factorized spatiotemporal attention to efficiently combine observations across cameras and time. Our video inference pipeline uses pose graph optimization to refine camera trajectories and assembles object tracks across input windows. We also develop a data annotation pipeline and release PAI-Foundation4D, a benchmark with dense depth and semantic occupancy labels for 1,253 driving clips. With benchmark-specific fine-tuning, Foundation4D achieves the lowest depth AbsRel on Waymo and nuScenes, the highest detection mAP on nuScenes, and the highest occupancy mIoU on OpenOccupancy among the compared methods. Our task-combination study shows benefits for both detection and geometry: co-training detection with geometry and occupancy improves top-class mAP by 1.2 points over detection-only training, while co-training geometry with detection reduces camera-to-rig translation error by $13.4\%$ relative to geometry-only training. We further use the predicted geometry and trajectories as priors for lidarless NuRec, supporting reconstruction of driving scenes for simulation.

World models enable embodied agents to anticipate action consequences, yet jointly modeling locomotion and manipulation across diverse embodiments remains challenging due to heterogeneous action spaces and limited multimodal supervision. We present Ropedia W0, the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. We curate Xp-Data, comprising 10k hours of egocentric human data and 3k hours of robot data with unified multimodal annotations and systematic quality control. A unified kinematic representation encodes locomotion through camera poses and manipulation through projected skeletons, with robot skeletons combining embodiment-specific forward kinematics with extensive manual keypoint annotation and correction. We also introduce Xp-Teleop, a long-horizon loco-manipulation benchmark based on high-quality dexterous humanoid teleoperation data. Evaluations demonstrate state-of-the-art performance across five cross-embodiment benchmarks (DreamDojo, $Ψ_0$, $ω$-HOME, Xp-Teleop, and AgiBot) and on held-out human egocentric video, from zero-shot inference to few-shot and full-data post-training. Scaling human pre-training data improves performance on both human and robot evaluations. After distillation, our model supports real-time, minute-long rollouts, enabling live VR interaction. Applications in policy evaluation, model-based planning, and world-action modeling further demonstrate the model's utility for embodied decision-making.

Transformers are widely assumed to exhibit token-decoding explainability. In particular, in vision transformer (ViT) classifiers, decoding individual patch tokens has become a common diagnostic tool due to its tendency to yield human-interpretable outputs. In light of the increasing popularity of this approach, an important question is whether token-decoding explainability is an inherent property of transformers that should be expected to hold across architectures, scales, and training objectives, or merely an artifact of some implicit bias. In this paper we systematically study token-decoding explainability across diverse ViT settings, spanning architectures, scales, and training objectives, using both controlled experiments and real-world models. We find that while some models indeed exhibit aligned and interpretable token decodings, others exhibit either inverted explainability maps or a complete lack of correlation with the image semantics. As we show, these tendencies toward different explainability behaviors are not intrinsic to the architecture and can vary arbitrarily with changes in the optimization landscape. We provide theoretical support for these observations by demonstrating a setting where a single fixed toy transformer architecture can provably implement the optimal classification rule in multiple ways, which differ in their patch-decoding explainability patterns. Our results show that token-decoding explainability is not an inherent property of trained transformers, but rather an outcome contingent on the problem setting and thus should not be a-priori assumed to hold. Code can be found in https://ronensc.github.io/Explainability-Illusion/ .

Preference optimization can improve every ranking metric while building a hidden, shared representation shift that the relative loss barely sees, one that can eventually destroy generation. We trace this failure to the common-mode offset, a shared directional shift in the model's hidden states that appears across token positions, prompts, and both chosen and rejected responses. Because it moves chosen and rejected representations together, the relative ranking loss barely reacts to its accumulation. As this offset grows, it projects onto the output layer to form a vocabulary well, elevating a single token's probability until it dominates decoding and causes generation collapse. The isolated offset alone reconstructs the vocabulary well, injecting the learned offset into an untreated base model recreates the collapse where norm-matched random directions do not, and masking this common component from the loss prevents the well in our flagship setting. Under the per-token ranking loss we study, the same collapse appears in 14 of 17 base-model lineages and under full-parameter fine-tuning. Moreover, the offset is not merely an artifact of collapse, since it is already present in textbook DPO and SimPO runs that form no well. Textbook DPO keeps it weakly aligned with any single vocabulary direction even with three times our base budget. Trained from base models for that budget, however, full-parameter SimPO drives it into collapse on both Llama-2 and Llama-3.1, and the same kind of offset grows, without a well, from a public SFT checkpoint under the official SimPO recipe. Across 17 UltraFeedback runs, the better a run ranks held-out responses, the less often two public reward models prefer its samples to the base model's. These findings expose a blind spot in standard ranking metrics and introduce a representation-level signal that can warn of generation collapse before it occurs.

Autonomous agents can increasingly propose ideas, modify code, and run experiments, creating an opportunity to accelerate scientific discovery using teams of collaborative AI researchers. To escape local minima in the discovery, we propose to encourage orthogonal exploration by constructing the team using agents with varying expertise. However, realizing this design requires more than diverse backgrounds. When experiments are costly, a team can test only a limited number of ideas, thereby creating the need to balance exploration and exploitation when selecting a subset of researchers to propose and execute new experiments. Furthermore, each result should shape what the team tries next. We therefore ask how AI researchers with different scientific backgrounds can build on one another's findings and sustain progress under a limited experimental budget. To address these challenges, we build a team of persistent AI researchers, each grounded in a different discipline, who keep private research histories, share experimental evidence, and receive research opportunities through adaptive allocation. We demonstrate the effectiveness of this design by deploying the team to search for linear attention architectures, where it outperforms GDN2 at 1.3B scale on language modeling, common-sense reasoning, and synthetic and real-world retrieval tasks. In controlled comparisons, removing any single organizational component from our design yields markedly weaker discoveries: without disciplinary backgrounds, the search plateaus early; without private histories, researchers make less progress despite identical shared evidence; and uniform selection falls short of adaptive allocation within the same budget. These results suggest that sustained autonomous discovery depends not only on stronger agents, but also on who explores, what they remember, and how research effort is allocated.

Long-context language models (LCLMs) power agents that act over long trajectories, from improving a codebase to running research experiments to forecasters that condition on lengthy histories. These applications draw on three distinct long-context capabilities: recall, where the answer is a fact present in the context; state tracking, where the outcome can be computed by performing a series of transformations to a specified state; and pattern learning, where the goal is to make a prediction by sampling from a distribution inferred from many observations. Existing benchmarks primarily focus on the first capability. In this work, we introduce LEAPS (Long-context Evaluation of Adaptive Pattern learning and State tracking), a collection of 12 novel datasets, spanning agentic workflows, behavioral modeling, and forecasting of market and world events. By construction, LEAPS is resistant to data contamination, extends to millions of input tokens, and uses reliable metrics. We evaluate 30 API and open models from 8K to 512K input tokens. LEAPS distinguishes models that perform near perfectly on recall tasks, finding that such models' state tracking and pattern learning abilities can vary significantly and degrade much more dramatically with increasing input lengths--even for pattern learning, where more observations should make learning a theoretically easier task. Moreover, thinking models often underperform their base counterparts, exhausting their generation budget or oversimplifying the task. We hope LEAPS can guide the development of the next generation of LCLMs.

Humanoid mobility over complex terrain requires accurate foot placement and coordinated whole-body motion capable of robust adaptation. Reinforcement learning (RL) policies trained without motion references generalize well across terrains, yet eliciting effective whole-body coordination through reward design and exploration alone remains challenging. In contrast, reference-based methods yield natural motions but struggle to adapt beyond the predefined trajectories in their training data. We introduce a learning framework that bridges these two approaches by using motion data strictly as a training prior. An explicit, self-supervised, attention-based contact planner predicts future body-environment contacts, while a contact-conditioned tracker translates these plans into coordinated whole-body motion. Together, they establish contact as an interface between sensory observations and control actions. This abstraction allows the policy to internalize the contact-aware structure of locomotion rather than directly replicating kinematics. Consequently, a single policy can exploit data-driven motion priors for natural whole-body coordination while retaining the flexibility to discover environment-dependent behaviors through reinforcement learning. Building on the learned contact planner and tracker, a subsequent fine-tuning stage bootstraps their acquired capabilities to explore environment variations beyond those covered by the original motion data. Operating entirely from onboard observations without any reference inputs at deployment, this framework enables versatile humanoid mobility in challenging environments.

Human video scales easily and does not need robot hardware, but existing methods require depth sensing, overlay the robot onto 2D video, or generate data only for the demonstrated objects. We present Real2Sem2Real (R2Sem2R), a real-to-semantic-to-real pipeline that turns one monocular RGB video of a human demonstration into thousands of robot demonstrations across the semantic object instances of a category. R2Sem2R tracks the SE(3) pose of each object with DINOv3-distilled Gaussian splats, outperforming existing methods and FoundationPose with estimated metric depth. R2Sem2R then aligns digital cousins, each generated from one or a few photographs, to the demonstrated objects with a semantic alignment. This alignment matches whole objects and the functional regions of target objects, such as a coffee machine's tray. R2Sem2R maps the human hand to a feasible robot grasp that extends to the cousins. Robot demonstrations are generated either in IsaacSim, for photorealistic rendering, or MuJoCo, for contact physics, and R2Sem2R can generate over 2,000 episodes in about two hours. On two real robots and four tasks, we train a vision-language-action policy on data from one human video per task, and policies trained on our data match or exceed policies trained on teleoperation across diverse object instances. The project page is available at https://real2sem2real.github.io/.

Despite recent advances in policy learning, long-horizon manipulation remains difficult because learned policies must avoid compounding errors while preserving future feasibility. While search-based planning can explicitly reason over such future consequences, it quickly becomes expensive in high-dimensional continuous action spaces. This raises a central question: how can long-horizon reasoning from search be transferred into policies in continuous action spaces? To this end, we propose DISCO: Distilled Tree Search over Continuous Actions. DISCO uses progressive-widening MCTS in physics simulation to generate long-horizon plans. During search, a learned policy proposes continuous actions, while a learned value function scores actions by their expected progress toward task completion. Using GPU-parallelized simulation, DISCO periodically distills MCTS-generated plans into the policy-value model, improving the action proposals and value estimates used in subsequent searches. Finally, the learned policy-value model can be directly deployed without the need for simulator-based search at test time. Across simulated and real-world rearrangement tasks, DISCO outperforms behavior cloning, reinforcement learning, and search-based baselines.

Video models provide rich priors for generalizable robotic manipulation, but plausible visual futures are not always executable. We introduce DreamCan, a hierarchical policy that grounds video models through sparse visual planning: a high-level planner adapted from a pre-trained video model and a low-level visuomotor policy. Rather than predicting dense trajectories, the high-level planner generates sparse visual subgoals that balance reachability and task utility. From offline, action-free robot videos, we estimate these quantities and use dynamic programming to select subgoal sequences without manual annotations. The resulting subgoals are more reliably attained than uniformly spaced alternatives, and the full system outperforms language, dense-video, and uniform-subgoal planners, reaching high planning success rates on Language-Table tasks and 91.2% success on LIBERO-Long with a 0.1B-parameter low-level policy. Additional suboptimal videos and test-time compute further improve success. The system also performs challenging real-world bimanual tasks, including box folding and packing with unseen object combinations. More results are available at https://dreamcan.github.io/.

Humanoid loco-manipulation is particularly challenging because it requires coordinating locomotion, balance, and object interaction, making it difficult to learn from limited demonstrations or from scratch through Reinforcement Learning (RL). We introduce DiPLO (self-distilled Diffusion Planning for Loco-manipulation), a framework that combines a self-distilled diffusion planner with a low-level whole-body tracking policy. Our planner is pretrained via supervised learning on motions derived from human-to-humanoid retargeting, while the tracker is trained via RL to execute them. DiPLO centers around two complementary post-training stages. First, we introduce Self-Distillation Diffusion Policy Optimisation (SD:DPO) to remove the planner's dependence on strong task conditioning, which we identify as a crucial ingredient for effective pretraining, as downstream fine-tuning does not recover comparable performance when it is omitted.. We then fine-tune this distilled model with Diffusion-Augmented Proximal Policy Optimization (DA:PPO), a maximum-entropy RL method for diffusion models, to improve task success and align generated plans with the frozen tracker's execution capabilities. Together, these stages enable motion generation through state-history and goal conditioning alone, enabling 99% task success in pick-carry-place tasks, compared to just 45% for the pretrained model. The resulting planner also generates motions for unseen object goals, with further evaluations on kicking and pushing/pulling tasks. Finally, we deploy our pipeline on the hardware, using our DA:PPO planner-tracker on a Unitree G1, achieving repeated successful executions without any real-world fine-tuning.

Consensus is a fundamental problem in distributed computing and at the core of state machine replication (SMR). Despite decades of active human research, discovering practical low-latency solutions to the problem is still open, as it requires combining protocol knowledge, correctness reasoning, implementation, and experimental evaluation. We report on how AI helped us discover A-greement, a new consensus protocol for geo-replicated state machines that significantly outperforms state-of-the-art solutions. Averaged across six deployments with 5 or 9 replicas, A-greement reduces the mean latency of a geo-distributed key-value store over the fastest baseline by 12.2% and 33.3% on write-heavy and read-heavy workloads, respectively, and matches the fastest baseline on read-only workloads. A-greement combines lease-authorized local reads, direct client learning for leader-ordered writes, and a fast fixed-quorum per-key write path, selected according to network topology and workload. At the heart of our discovery lies AI-greement, an agent harness for the "autonomous" discovery, implementation, experimentation, and verification of candidate protocols. AI-greement integrates (1) a concept-oriented protocol knowledge base, (2) a modular implementation skeleton, and (3) a Lean-based verification framework with reusable safety definitions, reference specifications, and proofs. AI-greement is of independent interest: its Lean formalization uncovers several previously unreported vulnerabilities in existing consensus protocols, and it provides guidelines on how domain-specific harnesses can turn general-purpose coding agents into effective tools for (distributed) computing research. Yet we argue that human consensus experts remain essential before, during, and after the search: agents reshape the researcher's role rather than end consensus research.